Why AI Rules Are Hard to Write Clearly
One AI model can help write a birthday invitation, screen a job application or support a medical workflow.
That flexibility makes AI useful. It also makes clear and durable rules unusually difficult to write.
Why one adaptable technology is difficult to define with fixed legal language.
The hardest part is not merely that AI changes quickly. The same model can become part of many systems with completely different consequences.
Imagine a new material that can be used in a toy, a car, a surgical instrument or a bridge.
A rule that says “this material is safe” would be too broad. A rule that says “this material is dangerous” would be equally unhelpful.
The sensible questions would concern the finished product:
- How is the material being used?
- What pressure will it face?
- What happens if it fails?
- Who is responsible for testing it?
General-purpose AI creates a similar problem.
A model is not always a complete product. It may be one component inside a much larger system.
One model can support many uses
A language model may be used to:
- rewrite an internal email
- summarise a meeting
- answer customer questions
- search medical records
- rank job candidates
- help investigate suspicious financial activity
The underlying model may be similar in each case. The risks are not.
A bad summary of an informal meeting may be annoying. A bad summary placed inside a medical workflow could affect a serious decision.
This makes it difficult to regulate only by model name, model size or technical architecture.
The model is only one part of the system
An AI system can include:
- a model
- instructions supplied by the developer
- documents retrieved from a database
- rules about who may use it
- software tools the model can call
- a user interface
- human review procedures
Changing any of these parts can change the system’s behaviour or risk.
A model that only drafts text is different from the same model connected to a tool that can approve payments. The second system has greater authority even if the underlying model has not changed.
Definitions must draw boundaries
Before a rule can be applied, it needs to identify what falls inside its scope.
That creates difficult boundary questions:
- What counts as an AI system?
- When is software only using AI as a minor feature?
- When does a general-purpose model become part of a high-risk use?
- Is the model provider responsible, the company using it or both?
- Does a small change create a new system version?
Definitions that are too broad can place heavy requirements on low-risk tools.
Definitions that are too narrow can leave consequential systems outside the rules.
Detailed rules can become outdated
A rule can describe an exact technical method.
For example, it could require one particular type of test or one named safety tool. That sounds clear, but the method may become outdated as technology changes.
The alternative is to write an outcome-based rule. It might require a system to reach an appropriate level of robustness or allow effective human oversight.
That wording lasts longer, but it creates a new question: what counts as appropriate or effective?
This is one reason standards, guidance and industry practices often sit beside formal rules. The law can state the required outcome while supporting documents explain methods that may help achieve it.
Probabilistic behaviour adds another difficulty
Many AI models generate outputs by selecting among possible continuations.
The exact answer can change when the prompt, context, settings or retrieved information changes. Even similar prompts can sometimes produce different results.
This does not make regulation impossible. Many regulated fields already manage uncertainty.
Medicine, finance, aviation and industrial safety all deal with systems that cannot guarantee perfect outcomes.
The challenge is to define acceptable processes, testing levels and safeguards when the behaviour cannot be reduced to one fixed answer.
Responsibility may be divided
Modern AI products often involve several organisations.
One company may train the model. Another may host it. A third may add documents, tools and instructions. A fourth may use the finished system with customers or employees.
When something goes wrong, each organisation may control a different part of the system.
Clear governance must therefore ask:
- Who chose the model?
- Who decided the intended use?
- Who tested the final system?
- Who monitors it after deployment?
- Who can stop or change it?
A rule that assigns all responsibility to one participant may not match how the system was actually built.
Why laws sometimes sound vague
Terms such as “reasonable,” “appropriate” and “sufficient” can feel unsatisfying.
They are sometimes used because one fixed technical threshold would not suit every system.
A customer-service assistant and a medical decision-support tool should not necessarily face identical testing requirements.
Flexible wording allows the required controls to increase with the risk. The cost is that organisations and regulators must interpret what the wording means in a particular case.
- One model can become part of many systems with different levels of risk.
- Detailed technical instructions can become outdated quickly.
- Broad outcome-based requirements need interpretation and evidence.
The plain-English takeaway
AI rules are not difficult only because lawmakers move slowly.
They are difficult because AI is adaptable. The same model can be placed in a harmless writing tool or a system that affects someone’s job, health or access to a service.
Good rules must remain flexible enough to survive technical change while being clear enough for organisations to follow and regulators to enforce.
Comments
Post a Comment